AI-Enhanced Structural Health Monitoring with a Multi-Rotor Aerial Vehicle

Roads, bridges, tunnels, railways and canals are crucial for the transportation of goods and people in modern society. Modern infrastructure is subject to damage and requires therefore regular inspection and maintenance to remain safe and functional. While this can be done manually, large parts of the mentioned infrastructure are hardly accessible, which is a main motivational factor for the use of multi-rotor aerial vehicles (MAV s) for inspection and maintenance tasks. MAV s, paired with novel deep learning computer vision techniques, such as instance segmentation, are emerging as an obvious choice for automating certain parts of structural health monitoring (SHM). In this manuscript, we provide an overview of available SHM datasets and how they can be utilized for the task of autonomous crack detection. We use available datasets to train and compare two neural network architectures for instance segmentation. The segmented crack instances are then localized in a global coordinate frame to perform autonomous mapping of the potentially dangerous infrastructure defects. Experimental studies demonstrate the effectiveness of the proposed SHM approach, which results in precisely localized infrastructure defects marked on the global map.